Fable-5-Qwen3.5-4B
Model Details
Model Description
Fable-5-Qwen3.5-4B is an instruction-tuned language model built on top of Qwen3.5-4B. It is fine-tuned using the Glint-Research/Fable-5-traces dataset, which consists of agent trajectories and reasoning traces designed to improve long-horizon planning, tool usage, and complex task execution.
The model is intended for research purposes involving agentic reasoning, software engineering, planning, and multi-step instruction following.
- Developed by: Glint Research
- Funded by: Not publicly disclosed
- Shared by: Glint Research
- Model type: Causal Language Model (Decoder-only Transformer)
- Language(s): Primarily English (multilingual capabilities inherited from Qwen3.5)
- License: Apache-2.0
- Finetuned from model: Qwen/Qwen3.5-4B
Model Sources
Uses
Direct Use
This model is suitable for:
- General chat
- Instruction following
- Multi-step reasoning
- Software engineering assistance
- Agent planning
- Tool calling workflows
- Code generation and debugging
- Research on reasoning traces
Downstream Use
Possible downstream applications include:
- Autonomous coding agents
- Research assistants
- Workflow automation
- Interactive tutoring systems
- Multi-agent systems
- Long-horizon planning tasks
Out-of-Scope Use
This model is not intended for:
- Medical diagnosis
- Legal advice
- Financial decision making
- Safety-critical autonomous systems
- High-risk domains requiring guaranteed correctness
Human verification is recommended whenever incorrect outputs could cause harm.
Bias, Risks, and Limitations
Like other large language models, Fable-5-Qwen3.5-4B may:
- Produce factually incorrect information ("hallucinations")
- Generate biased or culturally dependent responses
- Make reasoning errors despite producing convincing explanations
- Be sensitive to prompt wording
- Inherit limitations from its base model and fine-tuning data
Users should carefully validate outputs before deploying them in production environments.
Recommendations
For best performance:
- Use clear, structured prompts.
- Break complex problems into smaller subtasks.
- Verify generated code before execution.
- Keep a human in the loop for important decisions.
Getting Started
Example using Transformers:
1from transformers import AutoTokenizer, AutoModelForCausalLM
2
3model_name = "YOUR_MODEL_NAME"
4
5tokenizer = AutoTokenizer.from_pretrained(model_name)
6
7model = AutoModelForCausalLM.from_pretrained(
8 model_name,
9 torch_dtype="auto",
10 device_map="auto"
11)
12
13messages = [
14 {"role": "user", "content": "Explain how quicksort works."}
15]
16
17text = tokenizer.apply_chat_template(
18 messages,
19 tokenize=False,
20 add_generation_prompt=True
21)
22
23inputs = tokenizer(text, return_tensors="pt").to(model.device)
24
25outputs = model.generate(
26 **inputs,
27 max_new_tokens=512
28)
29
30print(tokenizer.decode(outputs[0], skip_special_tokens=True))
Training Details
Training Data
The model was fine-tuned using the Glint-Research/Fable-5-traces dataset, which contains synthetic reasoning traces and agent trajectories intended to improve multi-step planning and task execution.
Training Procedure
Preprocessing
Training data was formatted into conversational instruction-response pairs compatible with the Qwen chat template.
Training Hyperparameters
Detailed hyperparameters have not been publicly released.
Training likely followed standard supervised fine-tuning (SFT) procedures.
Evaluation
Testing Data
No standardized benchmark results have been publicly released.
Factors
Performance may vary depending on:
- Prompt quality
- Task complexity
- Context length
- Reasoning depth
- Tool availability
Metrics
No official evaluation metrics have been published.
Results
Qualitative evaluation indicates improved performance on:
- Long-horizon reasoning
- Tool-use planning
- Multi-step coding tasks
- Agent trajectory generation
However, quantitative benchmark results are not currently available.
Model Examination
The model inherits the multilingual capabilities and general knowledge of Qwen3.5 while specializing in reasoning and agent-oriented tasks through supervised fine-tuning.
Environmental Impact
Training infrastructure details have not been publicly disclosed.
- Hardware Type: Not publicly disclosed
- Hours used: Not publicly disclosed
- Cloud Provider: Not publicly disclosed
- Compute Region: Not publicly disclosed
- Carbon Emitted: Unknown
Technical Specifications
Model Architecture
- Architecture: Decoder-only Transformer
- Base Model: Qwen3.5-4B
- Parameter Count: Approximately 2B
- Context Length: Inherited from Qwen3.5 (refer to the base model documentation)
Compute Infrastructure
Hardware
Not publicly disclosed.
Software
Likely trained using the Hugging Face Transformers ecosystem together with PyTorch.
Citation
If you use this model, please cite both the Qwen base model and the Fable-5-traces dataset.
Example:
1@misc{qwen35,
2 title={Qwen3.5},
3 author={Alibaba Qwen Team},
4 year={2025}
5}
More Information
This model is intended primarily for research and experimentation in reasoning, planning, and agentic AI. Users should evaluate the model carefully before deployment in production systems.
Model Card Authors
Glint Research
Model Card Contact
Please refer to the model repository or organization page for contact information.
